Cyberthreats/Incidents

AI in 2026: the threat outside and within

Roman Reznikov

Roman Reznikov

Analyst, Research Group of PT Cyber Analytics

About the report

This report examines the evolution of AI-related threats and their impact on cyberattacks. It continues our series on AI technologies, following previous reports on the use of AI in cyberdefense and key trends in AI development. The report is based on Positive Technologies' internal expertise, incident investigation findings, and data from reputable external sources.

Because of reputational risks, most cyberattacks remain publicly undisclosed. As a consequence, even companies specializing in incident investigation and analysis of hacker activity are unable to quantify the precise number of threats. The purpose of this report is to draw the attention of businesses, cybersecurity professionals, and anyone interested in the current state of information security to the most relevant AI-driven threats and the defenses needed to address them.

This report considers each mass attack (for example, phishing emails sent to multiple addresses) as one incident, not several. For explanations of terms used in this report, please refer to the Positive Technologies glossary.

Summary

  • AI is a major force reshaping the cybersecurity landscape, affecting all of the main attack methods cybercriminals use.
  • While fully autonomous AI cyberattacks remain hypothetical, cybercriminals increasingly integrate AI into their operations.
  • Professional APT groups and skilled attackers benefit most from AI, using it to automate and accelerate specific attack stages.
  • Rather than building their own tools, attackers primarily exploit legitimate AI models and services, bypassing weak safeguards and restrictions.
  • AI accelerates vulnerability discovery, shrinking the window between a vulnerability disclosure and the appearance of a working exploit.
  • Cybercriminals routinely generate phishing content, deepfakes, and malware code fragments using AI.
  • Attackers now experiment with embedding AI modules directly into malware, downloading interface-recognition models onto victim devices or connecting to remote AI services to generate commands.
  • AI risks are not limited to cyberattacks—they also come from businesses adopting the technology insecurely. Shadow AI, AI-assisted development, and agentic systems create a new, high-risk attack surface.
  • Shadow AI—employees using unmanaged AI tools—opens the way for data leaks that are difficult to detect.
  • AI-assisted development increases insecure code volumes; while syntactically cleaner, the code retains serious security flaws.
  • Attackers target AI infrastructure to gain initial access, steal data, and hijack computing resources or paid model quotas.
  • Capable of initiating independent actions, AI agents create new security risks. Agent failures are already triggering security incidents, making their supply chains a prime target for hackers.
  • Traditional security weaknesses—weak credentials, exposed services, outdated software, and poor development practices—still enable most successful AI-driven attacks.

Introduction

Artificial intelligence is a major force reshaping the cyberthreat landscape in 2026. AI affects social engineering, vulnerability exploitation, and malware deployment. While AI has enormous potential to automate, scale, and complicate attacks, its real-world impact remains limited, changing cyberattacks gradually rather than overnight.

We also explore a second major threat: businesses hastily adopting AI without securing the technology or its underlying infrastructure. The risk is no longer hypothetical. Insecure AI systems embedded in business processes already disrupt operations and leak confidential data, making AI security a top priority for the coming years.

AI in the hands of cybercriminals

Levels of AI use in cyberattacks

In our previous research, we mapped out five levels of AI adoption across the attack lifecycle. Since 2024, the core principles of how criminals use AI have remained fundamentally unchanged. Instead, attackers are scaling up proven techniques, routinely generating malware and phishing content, and using AI to automate both vulnerability discovery and exploitation.

Despite these advancements, fully autonomous, end-to-end attacks remain out of reach for even the most sophisticated models. Research shows that in attacks against corporate environments, AI tools rarely progress past the initial exploitation phase without human operators stepping in to guide the process and handle the manual steps.

Figure 1. Levels of AI use in cyberattacks. Red indicates methods involving widespread use of the technology

On underground forums, threat actors are exploring post-compromise AI use, primarily to parse and analyze stolen data. However, the underground community does not actively discuss the offensive exploitation of AI. Instead, most conversations focus on finding uncensored LLMs or bypassing safety guardrails. In fact, "AI" is still mostly used as a marketing label for scams, which heavily outnumber actual AI-driven tools. In fact, "AI" is still mostly used as a marketing label for scams, which heavily outnumber actual AI-driven tools.

Who uses AI in cyberattacks

The industry's early fear—that AI would empower low-skilled script kiddies—has not materialized. Instead, advanced persistent threats (APTs) and cyberespionage groups are gaining the most significant advantages. Threat intelligence routinely observes cyberespies deploying AI-generated code, phishing lures, malware modules, and AI-optimized operational workflows. Skilled operators use AI to accelerate routine tasks, validate outputs, and bypass technical limitations. While AI will not turn a novice into a master hacker overnight, its integration into an attack chain now serves as a major indicator of a threat group's overall sophistication.

What AI is used in cyberattacks

Throughout 2023 and 2024, underground markets were flooded with LLMs marketed as purpose-built for cyberattacks. In reality, most of these were either outright scams or legitimate models bundled with automated jailbreaks designed to answer any question by bypassing safety filters. WormGPT exemplifies this trend. When it appeared in 2023, it was widely discussed on underground platforms and received extensive media coverage. As we noted earlier, real user feedback indicated that the tool was largely ineffective. The project quickly shut down, and the numerous clones and later versions that followed never attracted comparable attention. A February 2026 leak finally laid the project's inner workings bare: WormGPT lacked the promised built-in tools like Nmap or DNS scanners and was strictly limited to text responses. It operated simply as a Mistral-7B model paired with a RAG dataset of cybersecurity news summaries and a basic system prompt instructing the model that it was unrestricted and not participating in malicious acts. Out of more than 4,000 leaked user accounts, only 200 held active subscriptions.

Figure 2. WormGPT leak offered for sale

Despite the prevalence of these scams, dark web markets continue to advertise "unrestricted" LLMs. We expect this demand to persist regardless of past project failures, largely driven by criminals' distrust of legitimate AI providers and their fear of deanonymization. These concerns are highly justified, as companies like OpenAI regularly publish reports on disrupting specific threat actors who abuse their platforms. Yet, as the WormGPT leak proves, custom LLMs built by cybercriminals offer no safety guarantees for their users either.

Figure 3. Warning against using models due to potential deanonymization
Figure 4. Advertisement for an unrestricted GPT in February 2026
Figure 5. Advertisement for an unrestricted GPT in April 2026

Over time, offerings built around custom malicious LLMs are likely to give way to the emerging jailbreak-as-a-service model, which combines methods for bypassing the restrictions of legitimate models. Commercial AI offers a level of response quality that the underground simply cannot replicate without resources comparable to top-tier tech companies. Furthermore, despite developers' best efforts, bypassing these restrictions is still not overly difficult. Just as we predicted, criminal applications will continue to progress right alongside legitimate AI tools.

Legitimate models will remain the staple of the cybercriminal AI arsenal. This will only change if training high-quality custom models becomes drastically simpler and cheaper, prompting attackers to shift to their own solutions. Until then, attackers will continue exploiting top-tier models like ChatGPT and Claude Code, as demonstrated by numerous examples throughout this research.

Figure 6. Discussion of models used to write malicious code

MITRE ATT&CK heat map: AI today, tomorrow, and in the future

To track AI's evolving role in the threat landscape, we mapped its current and future viability against the MITRE ATT&CK framework.

  • The MITRE ATT&CK matrix is a knowledge base developed and supported by the MITRE corporation based on analysis of real APT attacks. The MITRE ATT&CK matrix includes description of tactics and techniques used by adversaries during attacks against corporate infrastructure.

     

    Tactics are the matrix columns. They describe the attacker's objective and divide the attack into stages. For example, the Initial Access tactic (TA0001: Initial Access) describes actions used to enter a network and gain a foothold for the next steps.

     

    Techniques are the elements inside those columns. They describe specific actions attackers take to achieve their goals. For example, the Initial Access tactic includes the Replication Through Removable Media technique (T1091: Replication Through Removable Media). In this technique, attackers try to reach isolated systems by using infected media, such as USB drives.

We updated the heat map from our previous study, while keeping the same methodology. The matrix shows how soon cybercriminals may be able to use AI for each tactic, technique, and sub-technique. AI demonstrates broad offensive potential and is eventually poised to impact 100% of tactics and over 62% of all known techniques.

Figure 7. MITRE ATT&CK heat map

Download the full MITRE ATT&CK heat matrix.

We divide all techniques and sub-techniques into five levels based on the potential time horizon for the appearance of AI technologies in them.

Maroon (confirmed use). Threat actors currently use AI in 10% of real-world attack techniques, doubling the 5% share observed in 2024. MITRE explicitly acknowledged this evolution in March 2026 by introducing T1683 Generate Content to track AI generation.

Red (near-term potential). Covering 26% of techniques, this category includes areas where researchers have already demonstrated practical AI applications and published proofs of concept (PoCs), such as T1567 Exfiltration Over Web Service.

Orange (foreseeable future). Making up 7% of techniques, these require attackers to overcome significant technical barriers. Despite the complexity, cybercriminals continue to expand their toolsets. They are also integrating more AI technologies into them. This is reflected in the gradual movement of techniques in this category into the red and burgundy zones of the matrix. There is little doubt that this trend will continue. In the future, attackers will at least test every technique where AI can provide an advantage in cyberattacks.

Yellow (theoretically possible). Accounting for 19% of techniques, these areas remain impractical for real-world operations given current technological limits. However, AI technology does not stand still. If AI takes a qualitative leap forward, attackers will inevitably use it to complicate, automate, and scale attacks here as well.

Gray (unjustified). In these scenarios, applying AI simply does not add meaningful value, or traditional methods remain vastly superior.

Vulnerability exploitation

Over the past two years, AI capabilities for finding and exploiting vulnerabilities have grown significantly. In 2024, we focused mainly on research findings and experimental results. AI tools are now widely used by penetration testers, CTF participants, and bug bounty hunters. Threat actors also use this technology in real-world attacks to both find and exploit vulnerabilities.

When used to identify known vulnerabilities, current AI solutions show mixed results across different vulnerability types. They perform best at finding issues in APIs and web applications. Over the last six months, bug bounty platforms have been so overwhelmed by low-quality AI-generated noise, forcing security teams to waste time triaging useless reports, that the curl project even ended its program in early 2026. Interestingly, Standoff Bug Bounty has avoided this issue, despite its hunters actively experimenting with AI.

Threat actors are also increasingly using AI in their attacks, including those against state targets. In late 2025 and early 2026, attackers successfully tricked Claude Code and GPT-4.1 into acting as automated penetration testers. This allowed them to breach several Mexican government agencies and exfiltrate hundreds of millions of confidential records. This case is an example of how broadly AI can be used to prepare for an attack, handling everything from information gathering and vulnerability hunting to writing the actual exploits. By accelerating the entire attack cycle—reconnaissance, vulnerability discovery, and exploit development—AI drastically compressed the timeline. Notably, standard security hygiene—such as timely patching, strong credentials, network segmentation, and endpoint threat detection—could have prevented these breaches.

Beyond general-purpose LLMs, open-source offensive frameworks like CyberStrikeAI and HexStrike AI are expanding attacker capabilities. HexStrike AI, a platform managing over 150 AI agents for scanning, exploitation, and persistence, currently receives high praise on underground forums.

Figure 8. Feedback on the use of HexStrike AI
Figure 9. Feedback on the use of HexStrike AI

Meanwhile, a financially motivated threat actor recently deployed CyberStrikeAI to compromise over 600 FortiGate devices. Amazon Threat Intelligence noted the attacker possessed only average skills, yet successfully used the AI to rapidly identify exposed ports and weak credentials—common security flaws. This incident shows why fixing basic security flaws is so important: increased automation makes them much easier to find and exploit. The threat of automated vulnerability hunting and low-complexity attacks potentially affects most companies. For example, as we noted in our study, 97% of penetration tests on Russian companies revealed weak password policies, and 80% had outdated perimeter software with known vulnerabilities.

AI can be used not only to detect known security flaws, but also to discover new, previously unknown vulnerabilities. This leads to the growth in disclosed threats, which, according to NIST, has reached 263% over the past five years. This could cause these databases to fragment and specialize in software specific to certain regions. As a result, tracking multiple sources, including CVE and local databases like BDU and PT), will become even more important. By generating PoC code and exploits faster, AI drastically shrinks the remediation window for defenders. Since organizations will find it impossible to patch every threat in time, vulnerability management must focus on asset prioritization and infrastructure knowledge—specifically, understanding which components lead to non-tolerable events.

Figure 10. Post about generating vulnerability PoCs with ChatGPT

Leading models current at the time of this publication, like Claude Mythos and Chat GPT 5.4 Cyber, possess highly advanced capabilities for both discovering and exploiting zero-days. The development of these capabilities makes thorough security assessment and comprehensive pre-release testing much more important. This includes testing products and systems with the help of AI tools. While developers restrict the most advanced offensive models to vetted security teams, the reality of the modern threat landscape—heavily shaped by political motives, where digital attacks often complement kinetic military strikes—means offensive AI will not be abandoned for moral reasons (and is already officially adopted by U.S. intelligence). Controlled access to advanced models cannot stop them from spreading. On the day of Claude Mythos's limited release, a group of enthusiasts managed to access and use the closed model undetected. Combined with other leaks and unauthorized access incidents, this proves that even the most powerful, restricted tools will eventually fall into malicious hands.

Social engineering

Social engineering remains one of the most common ways attackers use AI. Attackers increasingly rely on AI to prepare convincing phishing lures. According to Microsoft, recipients click links in AI-generated phishing emails in 54% of cases, which is 4.5 times more often than in phishing emails written by humans. AI removes the manual work from phishing. By pairing language models with autonomous AI agents, hackers can fully automate the process—from researching a target to holding a convincing back-and-forth conversation. This turns mass phishing campaigns into highly personalized "spear-phishing" attacks on a massive scale. HoxHunt demonstrated in 2025 that AI agents can execute spear-phishing at the level of a human professional, and the APT-35 group has already deployed this in the wild. Moving forward, these AI-driven, mass-targeted attacks will only increase. AI generation has even become a mandatory module in underground "phishing training" courses.

Figure 11. Phishing training course distributed on an underground marketplace

Determining the exact volume of AI-generated phishing is nearly impossible, as high-quality model output is indistinguishable from human-crafted text. Still, HoxHunt estimates that while AI-generated emails made up about 4% of phishing volume in 2025, this figure spiked to 40–50% during the holiday season. Moving forward, the proportion of AI phishing will fluctuate alongside mass campaigns, but the overall trend will be a steady increase.

Early 2026 research showed that ChatGPT's translation skills now rival entry- to mid-level human translators. While professionals still handle linguistic nuances better, these models are more than capable of localizing phishing lures, enabling cybercriminals to easily expand into new regions. For example, CrowdStrike reported Renaissance Spider using AI to translate lures in ClickFix attacks, while Google observed UNC1069 using the Gemini LLM to generate content in a language atypical for the group. In the coming years, we expect threat actors to expand into new regions, drawing more players into cyber conflicts around geopolitical hotspots.

Notably, threat actors are not just using AI to automate, translate, and tailor their lures; they are also using it for their own operational security. Cautious users on underground forums are now using LLMs to rewrite their posts to avoid exposing their identities. Ironically, research shows that these exact same models are highly effective at cross-referencing facts for deanonymization.

Figure 12. Use of AI to avoid deanonymization

As AI-driven social engineering expands, it creates a severe risk of mass cyberattacks aimed at public manipulation. While we previously reported how AI could trigger a bank run, other mass disinformation scenarios will inevitably emerge. These campaigns pose a unique danger during times of geopolitical tension.

Figure 13. Use of AI for mass public opinion manipulation

AI-driven social engineering is no longer limited to phishing emails. Code repositories have become a major staging ground, with cybercriminals using AI to mass-generate fake projects and convincing descriptions. We are also seeing more complex scenarios. In March 2026, Kaspersky reported on attackers abusing the Bubble platform to spin up web apps with clean URLs and hard-to-analyze code that easily evade security filters. This setup acts entirely as a stepping stone to automatically redirect victims to the final phishing site. Analysts believe this evasion tactic may already be in use on underground phishing platforms.

Phishing platforms

Even before AI entered the picture, phishing platforms and toolkits gave cybercriminals a way to automate attacks and helped less skilled actors compensate for their lack of expertise. We discussed their role in social engineering in an earlier report. Major phishing platforms are likely to receive AI-powered upgrades in the near future. Some are already moving in that direction. In April 2025, for example, Darcula added AI features for creating phishing data-collection forms and automatically translating them into different languages.

Figure 14. Phishing platform with an AI-powered message-editing feature
Figure 15. Advertisement for a scam call center service

At the same time, not every phishing tool that uses AI represents a real advance in cybercriminal capability. As with malicious LLMs such as WormGPT, AI can be a marketing label or even a way to draw attention to a scam. SpamGPT is a good example. Despite receiving broad publicity, it has not attracted positive feedback on underground forums. This suggests that it may be a project designed to deceive cybercriminals themselves.

Figure 16. SpamGPT offered for sale
Figure 17. Cybercriminals' feedback on the tool's poor performance

Deepfakes

We have discussed the growing role of deepfakes in social engineering attacks in several previous studies. Alongside other types of AI-generated phishing content, fake video and audio remain the largest and the only truly mass-scale use of AI in cyberattacks. The volume of deepfakes continues to rise. According to DeepStrike, the annual number of new deepfakes increased 16-fold between 2023 and 2025. Cybercriminals use deepfakes in many ways: to manipulate public opinion, support disinformation around major political events, create fraudulent advertising, impersonate relatives and friends in attacks on individuals, bypass Know Your Customer checks, and commit employment fraud to place insiders inside targeted organizations.

The widespread use of deepfakes has created a market for deepfake-as-a-service on underground platforms, which we examined in detail in an earlier report. Prices range from $20 to $200 per minute of generated video. Criminal providers also use many of the same marketing tactics as legitimate businesses: the price per minute falls as the total length increases, but rises for urgent orders or work requested on weekends. Underground deepfake-as-a-service providers charge more than legitimate AI video-generation platforms, where synthetic video typically costs around USD 10–20 per minute. The higher price on the dark web market reflects not only guarantees and a more personalized service, but also the nature of the output. Criminally produced deepfakes do not contain watermarks or the "imperfect by design" limitations that some commercial tools include for safety reasons.

Figure 18. Deepfake creation services offered

Deepfakes have become a standard social engineering tool for cybercriminals, used in both large-scale campaigns and targeted attacks. Their use is likely to keep growing over the next few years. More advanced threat actors will increasingly combine fake audio and video with other social engineering techniques in complex, multichannel attacks. We also expect the deepfake-as-a-service market to expand. Underground forums are likely to see more offers for custom fake audio and video, along with platforms for generating deepfakes. As we noted in our research on malware trends, the growing use of deepfakes may also help explain why spyware is increasingly adding audio-capture features. Some groups are already building attack chains where stolen personal information, voice recordings, and video materials are immediately reused in the next deepfake attack. With AI helping criminals process stolen data and generate realistic conversations involving multiple deepfake participants, skilled attackers could repeat and potentially surpass the most high-profile bank heist carried out with the help of deepfakes.

Figure 19. A malware feature implying deepfake use

The spread of convincing fake audio and video is becoming a serious problem. It will force organizations to rethink some identity-verification and KYC processes and, more broadly, how they verify information and decide what to trust.

AI as a phishing theme

Cybercriminals are interested in AI not only as a tool, but also as a theme for social engineering. In this report, we have shown several examples of LLMs, modules, and tools advertised as offensive AI that did not actually deliver the promised capabilities. AI-themed lures are not limited to underground forums. Cybercriminals also exploit the popularity of AI tools in large-scale malware campaigns. Mandiant Threat Defense tracked one such campaign for an extended period. The attackers promoted what they claimed was a video-generation tool, using social media ads and fake websites. In attacks targeting developers, AI specialists, and IT professionals, threat actors often use fake repositories. In February 2025, for example, experts at Positive Technologies discovered a malicious campaign in the PyPI package repository. The attackers exploited the hype around DeepSeek to distribute spyware packages aimed at developers who wanted to integrate the model into their projects.

Malware

Malicious code generation: from innovation to routine

In 2023 and 2024, as cybercriminals began using AI in attack preparation, code generation became an obvious task in which AI could be used, alongside phishing text generation. Just as legitimate developers adopted vibe coding and AI copilots, criminal developers began using AI to build malware. In only two years, AI-generated malicious code has become widespread. Researchers have found signs of LLM use, such as overly detailed comments, "generated at" timestamps, and other telltale markers, in the Punishing Owl stealer, backdoors used by Konni and MuddyWater, Hive0163 ransomware group, cryptominers, and even a RAT disguised as a software library. Many of these samples still contain clear markers of AI generation, suggesting that cybercriminals are not embarrassed to use AI and often make little effort to hide it. Today, analysts still treat AI-generated code in malware as a notable detail. In the near future, however, AI will become a standard tool in cyberattacks, and its traces in code will no longer stand out.

Figure 20. Post on an underground forum about the possibility of creating ransomware with AI

Attackers are no longer using LLMs just to write isolated scripts. They are increasingly applying them to build complete malware tools. VoidLink is a notable example. In early 2026, Check Point researchers reported that a single attacker likely built the malware in a relatively short time using AI tools. The attacker appears to have used AI not only to write code, but also to plan and manage the development process. This does not mean that anyone with access to a powerful AI tool can now build sophisticated malware. The workflow used in the VoidLink case still required significant technical skill. In the near future, we expect to see more complex malware developed with substantial AI assistance and deployed in real attacks. The most effective malware samples, however, are likely to come from experienced threat groups rather than low-skilled attackers.

Figure 21. Discussion about using LLMs to write malicious code
Figure 22. AI-generated banking trojan distributed on an underground marketplace

Cybercriminals sometimes rewrite existing malware instead of generating new tools from scratch. They use AI to create new versions of old samples, including public code examples and leaked malware. A more advanced scenario involves retrieving malicious files from VirusTotal. We have recently seen interest in this approach on underground platforms. In theory, attackers could collect malware hashes from open sources, obtain the matching files, and then use AI to rework them into new samples.

Figure 23. Downloading VirusTotal files services offered

AI inside malware

As expected, attackers are not only using AI to generate malware code. They are also experimenting with embedding AI into malware itself, either to support specific functions or to generate commands after infecting a victim's system. For now, malware with built-in AI modules or functionality remains extremely rare. Only a few examples have been found.

Figure 24. AI-powered malware detected from 2024 to Q1 2026

The most notable real-world example appeared in July 2025. LameHug used LLM-generated commands to perform reconnaissance and collect information. The malware generated those commands dynamically using prompts embedded in its code. LameHug accessed Qwen 2.5 through the public Hugging Face API. This not only proves that attackers are using legitimate models, but also that organizations need to monitor the use of AI services. PromptLock, a ransomware tool created by researchers, used a similar capability to generate malicious scripts.

In 2024 researchers discovered TesseractStealer, which used an AI-based optical recognition tool to process screenshots. By early 2026, several malware families had taken this concept of "smart vision" further. The Android.Phantom clicker trojan uses TensorFlow machine learning models to analyze browser screenshots and identify ad elements. Like TesseractStealer, it downloads local models to the victim's device and uses them for recognition. These tools are, of course, not comparable to the large LLMs used in other AI-enabled malware. Even so, they point to an emerging Bring Your Own AI pattern. The PromptSpy dropper, which also targets Android devices, takes a more advanced approach. Attackers used Gemini to process information about UI elements and return step-by-step instructions telling the malware how to interact with the interface. This allowed the malware to keep the infected app in the list of recently used applications. The AI module gives the malware more flexibility. It helps the malware adapt to interfaces that vary by device and OS version—a task that is difficult to automate with traditional scripts.

AI could also become a channel for C2 communication and data exfiltration. Check Point researchers have already shown how attackers could use Grok and Copilot for this purpose. This potential abuse of LLMs again highlights the need to monitor connections to external AI models. It also overlaps with a broader security problem: many organizations still lack visibility into approved AI tools in their infrastructure and Shadow AI.

To sum up, attackers still have a wide range of ways to exploit AI in malware. For now, however, only one area has seen significant adoption: malicious code generation. Over the next few years, attackers are unlikely to generate entirely new categories of threats using AI. Instead, they will focus on making AI-assisted code generation a routine part of attack preparation. They will also continue experimenting with AI-powered malware features in areas where flexibility and adaptation matter, such as searching for and analyzing victim data. For simpler tasks, a model can be downloaded directly to the victim's device. But when attackers use AI to generate commands or guide the attack, they still need larger models. Today, that usually means the malware has to access them remotely. As the technology improves, attackers may be able to delegate these more complex tasks to compact BYOAI modules. If organizations do not properly monitor access to APIs, cybercriminals may use this traffic to hide malicious activity, including adaptive malware behavior, C2 communication, and data exfiltration—topics we discuss below. Looking further ahead, as local AI models become more common in enterprise environments, we may also see advanced malware designed to abuse those models' resources.

After the attack

Attackers are not using AI only to build malware or run specific functions during an attack. Models can also help after a compromise by processing large volumes of stolen data, including unstructured data and files in different formats—a capability that is especially useful for ransomware operators. Leaks and stolen data dumps can run to tens or even hundreds of gigabytes. For attackers, being able to review this material quickly and extract the most sensitive confidential information is highly valuable. They can use it to put more pressure on the victim or to raise the price if they later sell the data.

Figure 25. Data exchange offers AI-powered data processing services

Cybercriminals have also started using AI not only to process data faster, but to deceive other members of underground communities. For example, they can add synthetic records to real stolen databases to make the datasets look larger and sell them at a higher price. If the synthetic data is convincing enough, attackers may go further and create fake databases from scratch. As we have noted in our dark web research, fraud is also common within cybercriminal communities. AI will certainly become not only a popular topic in these communities, but also a tool for scamming other criminals.

Figure 26. Posts offering payment for expanding sold databases

In the next stage, attackers are likely to build AI-based analysis of stolen data directly into malware, including spyware, ransomware, and wipers. This would let them move faster and make attacks more targeted and harder to detect. The damage, however, could remain high: the malware would focus on the most valuable data.

The threat within AI

Cybercriminals are still using AI selectively. They automate the stages of an attack where the technology has already proved useful, while continuing to test new techniques. At the same time, a different problem is becoming more urgent: risks that come from AI systems themselves. Companies are adopting AI much faster than they are adopting responsible AI practices in development and day-to-day operations. Microsoft reports that only about half of organizations have put safeguards around generative AI. This creates risk in two directions. If employees need AI faster than the company can provide approved tools, Shadow AI appears. If the company rolls out AI faster than its processes and security controls can mature, it ends up with weak oversight of model output, unsafe integrations, and poorly controlled deployments. Over the next few years, organizations will need to find a workable balance. They want the productivity gains AI can offer, but they also need to defend against new threats to agents, models, and the infrastructure those systems interact with.

Even though AI has entered a "trough of disillusionment,"1 investment remains strong. Gartner expects annual global AI spending to rise by almost 50%, reaching $2.5 trillion. We are seeing the same trend in cybersecurity. In 2025, venture funds, investment firms, strategic investors, and private investors put more than $7 billion into AI tools for cybersecurity. In 2024, that figure was nearly $4 billion. Most of the money went to projects that use AI to detect attacks—the area where AI has produced the clearest practical results in cybersecurity. Investment in securing AI itself moved in the opposite direction in 2025. Its share fell by 7 percentage points, and total funding dropped 17% year over year. This happened even as the need for AI security became more obvious, with several incidents involving AI systems, including agents. Once these threats started causing real damage, investor interest returned. By the first quarter of 2026, AI security had become one of the main investment themes.

1 The disillusionment phase is one of the stages in Gartner's technology Hype Cycle. Interest declines when implementation proves difficult and expert expectations are not met. Investments continue only if vendors improve their products to the satisfaction of early adopters. 

Figure 27. Investment in AI in cybersecurity

The shadow threat

Shadow AI, meaning employees using unapproved and unmanaged AI tools, is becoming a larger problem. Microsoft research shows that one in three employees uses this kind of AI. Many do it to work faster. In the process, they may underestimate the risks or choose to ignore them. Like other forms of Shadow IT, Shadow AI creates serious security risks. If IT and security teams do not know which tools employees are using, they cannot control where company data goes. These tools can also create new, poorly protected attack surfaces.

The most obvious risk is data leakage. If 13% of chatbot prompts contain sensitive information, including credentials, network details, and personal data, then exposure of confidential data becomes only a matter of time. According to IBM, 20% of organizations that suffered a data breach linked the incident to Shadow AI. For approved AI tools, the figure was 7 percentage points lower. Shadow AI also makes breaches harder to handle. As with other unmanaged technologies, incidents are harder to detect, investigate, and contain. Response takes longer, and the cost rises: on average, breaches involving Shadow AI cause about $200,000 more in damage.

This problem is unlikely to disappear soon. Companies cannot always roll out trusted AI tools quickly enough. Local models and in-house solutions often lack the required capacity or functionality. Personal AI assistants are spreading rapidly. And the human factor will continue to put security at risk. The only realistic way to address Shadow AI is to combine technical controls with organizational work, including employee training.

Figure 28. Access to the development network of an Abu Dhabi government LLM allegedly obtained via Shadow AI

The generated threat

Development

The global shift toward AI-assisted development is already reshaping the cyberthreat landscape. Veracode research shows that between 2023 and 2026, models reached 95% syntactic accuracy in the code they generate. Security, however, has seen little improvement. In about half of cases, generated codebases still contained known vulnerabilities. Models performed relatively well on some checks. They scored above 80% in tests for resistance to SQL injection (CWE-89) and the use of insecure cryptographic algorithms (CWE-327). But they struggled with more complicated issues, such as protection against XSS (CWE-80) and improper output neutralization for logs (CWE-117). Only about one in seven tests for these vulnerabilities was successful. Current models still handle security tasks poorly when those tasks require context rather than simple pattern matching. These problems stem from insecure training data, vulnerabilities in training datasets, and a broader neglect of secure development practices. For example, researchers found 12,000 hard-coded passwords and API keys in Common Crawl, a dataset used to train LLMs. Meanwhile, model development continues to prioritize performance, a goal that syntactic accuracy metrics suggest is being achieved, often at the expense of security. As a result, the problem is getting worse.

Organizations cannot afford to ignore insecure AI-generated code or treat it as normal. If they do, security incidents will increase, both because of cybercriminal activity and because of software errors. The longer companies leave this problem without proper attention and control, the longer the list of incidents will become. Vibe coding has already contributed to several damaging incidents. Examples include outages at Amazon, the loss of almost $2 million by the Moonwell lending protocol because of an oracle misconfiguration (DeFi system security was covered in our previous study), and a data leak at the Tea dating app. As these cases show, the threat affects both small companies and startups that are trying to ship products as quickly as possible, and large businesses that see vibe coding as a way to cut costs.

Security

LLMs work by predicting the most likely next tokens. Over time, this means they tend to give similar answers to the same prompts. That creates a real security problem. When teams use large language models to configure infrastructure, the results can become standardized. Standardization makes attacks easier. AI can also create template-like users and spread weak default values. This is already becoming an important factor in some large botnet campaigns. The same issue appears when people ask LLMs to generate strong passwords. A single password may look complex. But if the same prompt is used many times, the generated passwords often resemble one another and follow the same pattern. It is only a matter of time before LLM-generated passwords appear in brute-force dictionaries. The risk will keep growing because people tend to trust model output, and because vibe coding continues to spread. As a result, template configurations and credentials generated by AI will become more common. This will make relatively simple, highly automated attacks much easier to scale.

Infrastructure threats

As companies build AI into more processes, AI infrastructure is becoming a major attack surface. APIs, MCP, model weights, system prompts, and local models are now targets for both familiar attacks and new techniques.

More than 2,000 AI-related vulnerabilities were discovered in 2025, a third more than the year before. This growth rate is twice as high as the overall growth in disclosed CVEs. Some of these vulnerabilities affect AI solutions directly. Today, however, the biggest risks often sit not in the models themselves, but in the infrastructure around them. One example is the April data leak at the startup Mercor. The incident was linked to a supply chain attack on LiteLLM, a multifunctional gateway for AI agents. Attackers released two trojanized versions of the library, which then made their way into PyPI.

Figure 29. Access to an AI platform offered for sale

APIs are another major source of AI infrastructure risk. The growing use of AI is driving an increase in attacks on open and poorly protected APIs. Wallarm research shows that more than half of AI APIs are publicly accessible. At the same time, 89% lack strong authentication and are exposed to attacks aimed at abusing other users' tokens and plans or stealing data. The McKinsey case shows how serious this can become. CodeWall researchers used the consulting firm's AI platforms to gain access to a large volume of confidential data. Notably, an AI agent was also used to find and exploit the vulnerability. Leaked API keys can also lead to unauthorized model use. One developer in Mexico reported that attackers spent $82,000 worth of tokens in two days. Normally, the victim spent about $180 per month. LLMjacking has already moved into mass, industrialized cybercrime. Attackers now look for vulnerable AI endpoints and resell the access they obtain. When AI hosts are compromised, the risk is not limited to wasted compute or stolen tokens. Attackers may also access data from earlier chatbot conversations, move laterally through the infrastructure, and gain unauthorized access to file systems and databases. Securing AI endpoints is therefore only part of the task. Companies also need to monitor token spending. Some organizations already track token usage to measure employee productivity or control costs, but the same data can also help detect attacks.

AI systems are also becoming targets for another reason: attackers want information about the AI itself. The goal is not always to damage the company using the system or to make money directly. In some cases, the value lies in understanding how the system works. This interest is growing for two reasons. AI developers are locked in a constant technical race, and cybercriminals want to build more capable tools of their own. As a result, leaks involving new AI products attract significant attention. In April 2025, a repository containing dumped system prompts and internal materials from several companies became widely discussed. In March 2026, Anthropic materials were found in an unsecured storage location. More advanced attacks can involve collecting information about the model itself. One example is model extraction and distillation. Google reported attempts of this kind against Gemini in early 2026. Interest in leaks and unauthorized access to information about AI systems is unlikely to decline in the coming years. A leak of major model weights is also possible, although less likely. Developers have treated this information as especially sensitive since the LLaMA weights leak.

Figure 30. Post about a Llama data leak
Figure 31. Leaked source code of an AI product

Agent threat

Agents are now one of the key technologies behind AI automation. They mark a shift from AI that assists people to AI that can perform tasks on their behalf. Unlike traditional LLM-based assistants, agents can initiate actions, interact with infrastructure and data, and make limited decisions. That gives them far more utility, but also creates new security risks. We covered the history, design, and impact of agents on business and the cybersecurity landscape in detail in our study of emerging AI technologies. Many of the threats described there still remain potential rather than widespread. But there is little doubt that cybercriminals will target this new attack surface. If they compromise it, the opportunities could be significant. One of those predictions has already come true: attacks on the software supply chain for agents. Snyk research found that more than a third of all Skills, the packages that allow agents to use tools, APIs, or system resources, contain at least one security vulnerability. These issues range from hard-coded keys to embedded malware. Researchers have also found problems in another part of agent infrastructure: API routers. In one analysis of 428 routers, security researchers identified 29 that were infected, spyware-related, or capable of injecting malicious code.

Most agent-related incidents so far have not been driven by cybercriminals. They have been caused by the agents themselves. These cases show clearly how dangerous such systems can become when they behave incorrectly. Agent errors have already led to leaks of confidential data, database deletions, and even the deletion of emails from a user's mailbox. The last case involved OpenClaw, an agent that gained rapid popularity in early 2026. OpenClaw is one of the clearest examples of an immature agent ecosystem. Early versions worked through a gateway exposed to the public internet. Researchers found numerous security issues and critical vulnerabilities in the tool, including a one-click remote code execution flaw tracked as (CVE-2026-25253). The risks were not limited to the agent itself. OpenClaw's open Skills marketplace also contained many malicious Skills. Over time, developers are finding and fixing these vulnerabilities. Newer and more secure versions are also appearing, including NVIDIA's NemoClaw. But the OpenClaw case shows how quickly raw, unsafe AI tools can spread. In a short time, they can put tens of thousands of systems at risk.

Organizations should not avoid agentic AI simply because it introduces new risks. But they should deploy it deliberately, with clear controls, limited privileges, and a realistic understanding of what can go wrong. AI agents can be highly effective, but they should be treated like any other powerful system: useful, but capable of causing damage if poorly configured or left unchecked. A secure rollout should start with a few basic questions.

1.       What can the agent access?

Apply the principle of least privilege. Give the agent only the access it needs to perform its assigned tasks, and nothing more. Where possible, run agents in isolated environments.

2.      How is the agent allowed to complete its tasks?

It is not enough to define the desired outcome. Teams also need to define acceptable ways to reach it. Without clear boundaries, an agent may take unsafe shortcuts and bypass controls.

3.      How is the agent's activity logged?

Teams need enough visibility to understand what the agent did, why it did it, what data it accessed, and which tools it used. Logging these actions helps teams investigate failures, tune controls, and prevent the same issue from recurring.

4.      What stops the agent when something goes wrong?

Safeguards should be in place before deployment. The goal is to stop unsafe behavior quickly and reduce the impact of errors.

In the next few years, another important security issue may emerge if self-evolving agents become more widely used: tracking how an agent changes over time. Research shows that self-evolving agents can degrade as they develop. If an agent repeatedly performs simple, routine tasks, it may gradually lose the ability to use tools effectively in more complex cases. This problem could be especially serious in security, where systems often process a constant stream of normal activity and only occasionally face bursts of malicious behavior. Degradation is not the only risk. Hallucinations and errors can also accumulate in an agent's memory. Over time, this can lead to mis-evolution, where the agent starts handling tasks incorrectly or ignoring security restrictions.

Conclusion

The impact of artificial intelligence on cyberthreats can no longer be ignored. Both extremes are dangerous: underestimating what the technology can do, and trying too early to defend against attacks that do not yet exist, while neglecting the security measures that are needed today. AI adoption by cybercriminals for offensive operations, and by companies for process automation, does not remove the need for basic information security. On the contrary, it makes that need even clearer. Many AI-related incidents have happened because organizations lacked long-established security controls. The root causes were familiar: weak credentials, outdated software, and poor secure development practices. AI automates and scales the exploitation of existing vulnerabilities. It also becomes a catalyst for new risks inside the infrastructure when companies deploy it too quickly, without enough maturity or protection. As threats grow and accelerate, organizations need a systematic approach to effective cybersecurity. They must assess new technologies realistically, protect their weak points during deployment, and use their strengths to detect and prevent attacks earlier.

Unlike cybercriminals, defenders face certain limits in how they can use AI. Even so, the defensive side is adopting the technology in a more systematic and continuous way. The key areas of cyberdefense automation and development covered in our previous study continue to mature and are being integrated into an increasing number of security solutions. These areas are attack detection, including the detection of unknown attacks, faster response, and support for security specialists. Benchmarks for evaluating large language models in security tasks show that current LLMs cannot fully take over attack detection. But they do not need to. Positive Technologies' ML specialists discussed these benchmarks in detail in a separate article. As we have emphasized many times, the role of AI in cybersecurity today is to complement and strengthen security tools. The technology works best when used together with established protection mechanisms. For example, in PT X, the ML module helps detect anomalies in incoming data and identify new attack patterns, supplementing classic detection methods.